The Insane Engineering Behind DeepSeek Harness DeepSeek Harness launched on August 13, 2026, and within 48 hours had picked up nearly four times its opening-day GitHub stars, a pace DeepSeek's own AI models never came close to matching. The reason has almost nothing to do with the model it runs on. DeepSeek Harness is built entirely out of interchangeable plugins (everything is a plugin), letting developers swap the AI model, the toolset, or the interface without touching the underlying code. An approach most AI coding agents avoid because it's notoriously hard to get right. This video breaks down the engineering that makes it work, including Cordis, the plugin framework underneath it that turns out to have a longer and more interesting history than the launch coverage let on. At the center of it is a deceptively simple problem: how do you let a system rebuild itself out of removable parts without it slowly falling apart from leftover, half-removed pieces? The answer involves two mechanisms working together: one that gives every action its own built-in undo instruction, and one that lets components adapt automatically when what they depend on changes. Whether this architecture becomes the standard other AI agent harnesses get compared against, or gets overtaken by something built the same way, is still very much an open question. - @thom.code